Publication Abstract

Marine spatial planning and conservation both require sufficiently detailed knowledge on seabed sediments and habitats. Yet, only a few countries have so far initiated and executed large-scale seabed mapping programmes due to the involved high costs. An alternative or complementary approach in the interim would make best use of data that is currently available. Typical data sets for such a task encompass seabed sediment data, bathymetry and modelled oceanographic data. Here we describe an approach that uses such data sets from the North Sea to map seabed sediment composition and EUNIS habitats through geo-statistics. The EUNIS habitat classification system uses substrate type, energy level and photic zone to discriminate among habitat classes in the upper three levels of the classification hierarchy.

We applied a hybrid spatial prediction model to map the sediment composition of the North Sea. The approach employed both correlation with auxiliary predictors and spatial autocorrelation in prediction. The % Sand, % Mud and % Gravel values were then analysed in a GIS to classify the seabed according to EUNIS level 3 habitats.

We applied a regression-kriging approach to map the variability of the dependant variable. The predictors included bathymetry at varying resolutions, derivatives from the bathymetric digital elevation model including: slope, rugosity, curvature, aspect and the bathymetric position index. Predictors derived from oceanographic models included seabed shear stress from tides and wave base.

We created a semivariogram of the residuals from the regression model to examine its spatial structure. As spatial auto-correlation was present, simple kriging was applied to interpolate the residuals between sample points. Finally the two layers were summed to produce the final prediction layer.

There are several benefits of this coupled approach where deterministic (regression modelling) and stochastic (kriging) elements are combined, as the use of the deterministic model alone would ignore the spatial auto-correlation present in the data and using spatial interpolation alone ignores underlying trends in the data that can be explained (at least partially) by using secondary variables.

This approach also allows for a range of deterministic models such as General Additive Models or Neural Networks to be applied to explain more complex non-linear relationships between the predictors and dependant variable, the residuals of which can then be tested for spatial autocorrelation.

The results are a detailed map of predicted sediment composition of the North Sea and an associated confidence layer.  The sediment composition is then applied to classify habitat types.